docs(napkin-math): prioritize compress variance follow-up

[neoneye]

Summary

• update the 2026-05-20 napkin_math plan now that PR napkin-math: Phase 2 extract — source-arithmetic preservation, threshold-pairing parity, source_text truncation #740 is merged
• mark Phase 2 extract prompt-side work as done on main
• reorder follow-up priorities around output-quality impact: compress variance first, source-preservation audit second, behavioural validation third, prompt hygiene fourth

Rationale

The clearest unresolved napkin_math quality issue is upstream signal loss in compression: the Paperclip OPC UA / p99 latency tripwires are absent from the v50/v51 digest, so extract cannot recover them. Source-preservation audit remains important, but it classifies loss rather than restoring missing compress signal.

Validation

• git diff --check